Charging and discharging control method and system for multi-user energy storage power station based on electricity price

Abstract

The invention discloses a multi-user energy storage power station charging and discharging control method and system based on electricity price, and the method comprises the steps: arranging a monitoring terminal at energy storage equipment, and collecting the load data of each energy storage equipment; carrying out data preprocessing on the collected load data; constructing a cascade neural network prediction model, training the model by using the preprocessed load data, and generating a regional total load demand prediction value in a future time period; designing a deep reinforcement learning model DRL by using the regional total load demand predicted value to obtain a preliminary power adjustment amount; designing a target function considering the battery aging cost by using the initial power adjustment amount, and determining an initial value range of a coefficient in the target function of the optimization model according to a constraint condition; and a multi-objective optimization problem is considered, a Pareto optimal solution set is obtained based on the initial value range, and a final power distribution scheme is determined. The method can accurately predict the load, optimize the charging and discharging strategy, reduce the power consumption and operation cost, balance the service life of the battery and the economic benefit, and improve the intelligent and efficient level of the energy storage power station.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system energy storage control, and in particular to a multi-user energy storage power station charging and discharging control and system based on electricity price. Background Art

[0002] In the context of global energy transformation, the application of energy storage technology has gradually become an indispensable part of the power system, especially in the case of large-scale access to renewable energy. With the increase in the proportion of unstable energy sources such as wind and solar energy, the power grid needs a more flexible and efficient energy management system to balance power supply and demand and ensure the stable operation of the power grid. Therefore, the charging and discharging control technology of energy storage power stations has become one of the focuses of power system research.

[0003] In the prior art, for example, the patent application with publication number CN119742837A adopts a load forecast combined with charge and discharge strategy optimization to adjust the charge and discharge plan of the energy storage power station according to the change of grid load. This method optimizes the charging and discharging behavior of the energy storage power station by predicting the grid load and combining information such as real-time electricity prices. This method relies on static rules and does not consider the battery health status and battery degradation in actual operation, resulting in accelerated battery aging under long-term operation, which may affect the economy and stability of the system. For example, the patent application with publication number CN119154358A constructs an objective function and introduces auxiliary variables to represent the absolute value of the charge and discharge amount. The objective function includes constraints such as the maximum capacity of the battery, the charging efficiency, and the number of charge and discharge times, and uses a linear programming algorithm to solve the charge and discharge amount of the energy storage power station within the target cycle. This method does not consider the health status of the battery, which may lead to overcharging and discharging of the battery and shorten the battery life. However, the calculation model of the linear programming algorithm in this method has certain limitations when dealing with large-scale and complex energy storage systems, especially when considering nonlinear factors such as battery health and aging, the optimization effect is not ideal. For example, the patent application with publication number CN119029881A sets upper and lower limits and physical constraints to ensure smooth and continuous changes in the amount of electricity in the energy storage power station during operation. Another example is the patent application with publication number CN108173275A, which is adjusted based on the relationship between the power generation price on the grid side and the response amount of electricity on the user side. The method updates the electricity price and the response of electricity through the grid-side benefit optimization function, and finally adjusts the charging and discharging plan on the grid side according to the response amount of different users to achieve coordination between the grid and the users. However, the power allocation of this method is carried out under static constraints, lacks dynamic adjustment capabilities, and fails to adjust according to real-time changes in battery health.

[0004] In summary, the existing power management methods have many limitations. For example, they do not fully consider the superimposed impact of dynamic fluctuations in multi-user loads on the power grid, resulting in inaccurate prediction and regulation of power grid loads; fixed electricity price strategies are difficult to match the real-time price fluctuations in the power market, and cannot effectively guide users to use electricity rationally and optimize market resource allocation; centralized control modes are difficult to achieve coordinated optimization when facing multiple energy storage devices, limiting the overall performance of the energy storage system; in addition, there is a lack of real-time closed-loop control of the operating status of mobile energy storage devices, which cannot ensure that mobile energy storage devices participate in power grid operations efficiently and safely. Therefore, there is an urgent need for a multi-user energy storage power station charging and discharging control technology that can monitor the health status of energy storage devices in real time and dynamically adjust power distribution to optimize battery life, improve power grid stability, and meet multi-objective optimization requirements. Summary of the invention

[0005] In order to solve the deficiencies in the prior art, the present invention provides a multi-user energy storage power station charge and discharge control method and system based on electricity price, which realizes real-time closed-loop control of battery status by introducing multi-objective optimization, deep reinforcement learning, cascade neural network and battery health assessment model, and can dynamically adjust and optimize according to the real-time health status of the battery. This not only solves the health management problem in traditional methods, but also greatly improves the flexibility and efficiency of charge and discharge control of energy storage power stations.

[0006] The present invention adopts the following technical solution.

[0007] The present invention proposes a charging and discharging control method for a multi-user energy storage power station based on electricity price, comprising: S1, collects the load data of each energy storage device, including the temperature, voltage and current of the battery, and performs preprocessing; S2, constructing a cascade neural network prediction model, using the preprocessed load data to train the cascade neural network prediction model, and outputting the regional total load demand prediction value; S3, based on the deep reinforcement learning model DRL, builds a power adjustment prediction model. Through the collected real-time electricity price signals, the current state of charge SOC of the energy storage device, the grid frequency regulation demand level, and the regional total load demand forecast value, the training process of the power adjustment prediction model is controlled in real time to obtain the preliminary power adjustment value of each energy storage device. ,in i =1,...,N, where N is the total number of energy storage devices; S4, using the initial power adjustment Construct an objective function that takes into account the battery aging cost, and determine the power adjustment penalty coefficient in the objective function of the optimization model based on the constraints. and the state of charge weight factor The initial value range of S5, power-based penalty coefficient adjustment and the state of charge weight factor The initial value range of and The optimal coefficient solution is obtained, and the final power allocation plan is determined according to the optimal coefficient solution; the charging and discharging control of the multi-user energy storage power station is carried out based on the power allocation plan.

[0008] Furthermore, in S2, the cascade neural network prediction model includes a temporal convolutional network TCN and a graph long short-term memory network Graph-LSTM; Collect user type characteristics and environmental parameters, input historical load data, user type characteristics and environmental parameters into the cascade neural network prediction model, and output the predicted value of regional total load demand.

[0009] Furthermore, S3 includes: S301, the state space includes the real-time electricity price signal, the current state of charge SOC of the energy storage device, the grid frequency regulation demand level and the regional total load demand forecast value, and the load forecast deviation coefficient obtained by calculating the deviation between the regional total load demand forecast value and the actual value and normalizing it; the state space is encoded and converted into the corresponding data form; S302, the regional total load demand forecast value with the set deviation is used as the action space constraint, the total power sum( ) The balance formula is: sum(Δ P i ) = P req ±γ%; in, For the i The initial power adjustment of the energy storage equipment, including i =1,...,N, where N is the total number of energy storage devices, P req is the forecast value of regional total load demand, γ% represents the setting deviation; S303, training strategy design, including distributed parallel training, priority experience replay PER and training termination condition setting.

[0010] In S301, the real-time electricity price signal, the state of charge SOC of the energy storage device, the grid frequency regulation demand level and the regional total load demand forecast value, as well as the load forecast deviation coefficient obtained by calculating the deviation between the regional total load demand forecast value and the actual value, are encoded, and the real-time electricity price is normalized to [0,1]; the current state of charge SOC of each energy storage device is converted into a 20-dimensional vector, each dimensional vector represents a different attribute of the SOC data.

[0011] Furthermore, in S3, the training process of the power adjustment prediction model is controlled in real time, including the operation cycle design: In the long period of minutes, the cascade neural network prediction model is updated and the future load curve with minute granularity is generated; In a short period of seconds, the DRL model generates power allocation instructions based on the latest electricity price and SOC status, and the quadratic optimization solver adjusts the initial power adjustment of each energy storage device. , satisfying the device constraints, where i =1,...,N, where N is the total number of energy storage devices; It also includes instruction issuance and execution and exception handling mechanisms.

[0012] Furthermore, the exception handling mechanism includes SOC equalization trigger conditions and emergency shutdown scenarios; The SOC balancing trigger condition means that when the maximum SOC difference in the energy storage group is greater than the set SOC difference threshold, the balancing charging mode is started; the maximum SOC difference is the difference between the maximum and minimum SOC values ​​of all energy storage devices in the same energy storage cluster; the balancing charging mode is to set the SOC middle value, discharge or charge the energy storage devices above or below the SOC middle value respectively, and limit the maximum power of the balancing process to a fixed multiple of the rated capacity of the battery in the energy storage device; The emergency shutdown scenario means that when the temperature of the battery in the energy storage device is greater than the set temperature threshold or the voltage change rate in the two seconds before and after is greater than the change rate threshold, the charging and discharging circuit is cut off, the liquid cooling system is started, and the cloud is reported to trigger the backup power supply switch.

[0013] Further, in S4, the initial power adjustment amount Based on the Q-learning quadratic optimization algorithm, an objective function model considering the battery aging cost is established: Aging cost = ; in, is the power adjustment penalty coefficient, For the i The initial power adjustment of the energy storage equipment, including i =1,...,N, where N is the total number of energy storage devices, is the state of charge weight coefficient, SOC i For the i The state of charge of the energy storage device, SOC ref Indicates the initial capacity or benchmark capacity of the energy storage device under standard test conditions; The aging costs need to be kept to a minimum, namely: .

[0014] Furthermore, set constraints: ; in, SOC i express t +1 moment i The state of charge of the energy storage device, i =1,...,N, where N is the total number of energy storage devices; Indicates i The initial power adjustment of the energy storage equipment; Indicates i The initial power of the energy storage equipment; and Respectively represent i The minimum and maximum values ​​of the power of each energy storage device after adjustment; Represents the total power of N energy storage devices; According to the constraints, the penalty coefficient is adjusted through the interaction between the Q-learning agent and the environment. and the state of charge weight factor The initial value range of .

[0015] Furthermore, the multi-objective optimization model considers multi-objective optimization problems, which include minimizing grid fluctuations, maximizing economic benefits, and minimizing battery aging; the solution steps of the multi-objective optimization model include weighted space sampling method, non-dominated sorting, and decision maker selection mechanism; The steps of weighted space sampling method are: Penalty factor for power adjustment and the state of charge weight factor Perform grid sampling to obtain power allocation solutions under different parameter combinations and generate candidate sets including multiple groups of optimization solutions; The steps of non-dominated sorting are: Determine a first frontier surface according to the candidate set, that is, a solution set that is not dominated by other solutions; The steps for decision makers to choose a mechanism are: Provides two selection modes: economic priority and life priority, and selects the appropriate solution from the solution set that is not dominated by other solutions according to actual needs.

[0016] Furthermore, it also includes S6, calculating the health of the energy storage device, and performing closed-loop control on the weight parameter update of the power adjustment prediction model and the multi-objective optimization model; the formula of the equipment health evaluation model is as follows: ; in, Indicates i The health scores of energy storage devices are i =1,...,N, where N is the total number of energy storage devices, ; SOH indicates the battery health status, ,in, Indicates the current capacity. Indicates the initial capacity; It represents the temperature influence coefficient, and the calculation formula is: , T is the average battery temperature, T opt For the best working temperature, k is a constant; Indicates the cumulative number of equivalent cycles; Indicates the battery design life; , and They represent the battery health status weight, temperature impact weight and cycle number weight respectively.

[0017] The present invention also proposes a multi-user energy storage power station charging and discharging control system based on coordinated optimization of electricity price and electricity quantity, including a data acquisition and preprocessing module, a load prediction module, a preliminary power adjustment amount calculation module, a power adjustment amount optimization module and a weight parameter update module: The data collection and preprocessing module collects multi-user load data on the user side and performs data preprocessing; The load forecasting module inputs multi-user load data into the cascade neural network forecasting model to generate load demand forecast values ​​for future periods; The preliminary power adjustment calculation module builds a deep reinforcement learning model DRL and outputs the preliminary power adjustment of each energy storage device; The power adjustment optimization module establishes a secondary optimization model that takes into account the battery aging cost and determines the initial value range of the power adjustment penalty coefficient and the state of charge weight coefficient in the objective function of the optimization model; The optimal power allocation scheme design module obtains the optimal coefficient solution based on the preliminary value range and determines the final power allocation scheme; The closed-loop control module calculates the health of the energy storage equipment and performs closed-loop control on the weight parameter update of the power adjustment prediction model and the multi-objective optimization model.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention integrates the load time-space correlation analysis into the electricity price response mechanism. The existing technology usually only considers the load characteristics or electricity price factors separately, and rarely integrates the load time-space correlation analysis into the electricity price response mechanism to guide the operation of energy storage power stations. The load time-space correlation analysis is integrated into the electricity price response mechanism to combine the time-space variation characteristics of the load in different time periods and different regions with the real-time electricity price fluctuations, providing a more accurate basis for the charging and discharging arrangement of energy storage equipment. It reduces the electricity cost, improves the regulation ability and enhances the stability.

[0019] 2. The battery life-economy joint optimization objective function designed by the present invention comprehensively considers the battery aging cost and economic benefits. When formulating charging and discharging strategies, the existing technology often focuses on a single goal, either simply pursuing economic benefits while ignoring battery life, or over-focusing on battery life at the expense of economic benefits. The present invention designs a joint optimization objective function that comprehensively considers the battery aging cost and economic benefits, and balances the relationship between battery life and economic benefits. It extends the battery life, reduces operating costs and improves benefits.

[0020] 3. The operation of energy storage equipment in the prior art may lack a complete closed-loop control system, and there are deficiencies in load forecasting, decision-making and execution feedback, making it difficult to achieve rapid response and efficient operation. The present invention realizes the "forecast-decision-execution" closed-loop control of mobile energy storage equipment groups, making the operation of energy storage power stations more intelligent and efficient. A closed-loop control system is constructed from accurate load forecasting, to scientific decision-making based on forecast results and real-time equipment status, and then to quickly adjust the equipment operation status according to decision instructions and monitor feedback in real time. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a method flow chart of a method for controlling charging and discharging of a multi-user energy storage power station based on electricity price of the present invention; Figure 2 It is a schematic diagram of the training steps of the deep reinforcement learning model DRL of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only embodiments of a part of the present invention, rather than all embodiments. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the protection scope of the present invention.

[0023] The present invention proposes a charging and discharging control method for a multi-user energy storage power station based on electricity price, such as Figure 1 This embodiment takes an industrial park energy storage power station (including 20 mobile energy storage devices) as an example to explain the implementation steps of each link in detail.

[0024] S1, deploy the hardware system, including the data collection layer, edge computing layer and cloud training platform.

[0025] In the data collection layer, smart meters are deployed and installed on the user side (factories, commercial buildings, and residential areas) to collect 15-minute load data (with an accuracy of ±0.5%). Monitoring terminals are set up on energy storage devices to collect battery temperature, voltage, and current.

[0026] Specifically, in the edge computing layer, a local controller is configured to perform load prediction model reasoning and a regional edge server is configured.

[0027] Furthermore, data preprocessing is performed on the collected load data.

[0028] Specifically, the load data preprocessing step includes missing value processing and feature engineering. In the missing value processing step, linear interpolation is used to fill short-term missing values ​​(<5 minutes), and long-term missing values ​​(>30 minutes) trigger the data source verification alarm; in the feature engineering step, the load data is standardized, including Z-score normalization of continuous variables such as temperature and load, and the construction of a spatiotemporal feature matrix.

[0029] S2, construct a cascade neural network prediction model, use the preprocessed load data to train the cascade neural network prediction model, and generate the regional total load demand prediction value for a fixed period in the future.

[0030] Specifically, the cascade neural network prediction model includes the time convolution network TCN and the graph long short-term memory network Graph-LSTM. TCN uses dilated causal convolution to extract load time series features, and the specific structure includes convolution layer and activation function. Among them, the convolution layer includes 3 layers of dilated causal convolution (convolution kernel size = 8, dilation factor = 2), and the activation function is Gaussian error linear unit GELU. Graph-LSTM uses the dynamic time warping (Dynamic Time Warping, DTW) algorithm to calculate the similarity of user load curves and construct an adjacency matrix.

[0031] The cascade neural network prediction model is trained on the cloud training platform.

[0032] Specifically, the preprocessed historical 90-day load data is divided into 70% for training, 15% for verification, and 15% for testing. During the training process, the cascade neural network prediction model is input with historical load data, user type characteristics, and environmental parameters. The model outputs the regional total load demand forecast value at the granularity level of the next 15 minutes. The regional total load demand forecast value is verified using the verification set and judged using performance indicators, including the mean absolute percentage error MAPE. In this embodiment, when MAPE<8%, it is considered that the prediction accuracy of the model can meet the use standard.

[0033] S3 collects real-time electricity price signals, the current state of charge (SOC) of energy storage devices, and the grid frequency regulation demand level. It uses the regional total load demand forecast value to build a power adjustment prediction model based on the deep reinforcement learning model DRL. It controls the training process of the power adjustment prediction model in real time to obtain the preliminary power adjustment value of each energy storage device. ,in i =1, ..., N, where N is the total number of energy storage devices.

[0034] In deep reinforcement learning, for the regional total load demand forecasting problem, the current or future load demand forecast value is an important component of the state because it directly reflects the operating status and demand trend of the power system.

[0035] Specifically, the state space of DRL includes real-time electricity price signals, the state of charge (SOC) of energy storage devices, the grid frequency regulation demand level, and the regional total load demand forecast value, as well as the load forecast deviation coefficient obtained by calculating the deviation between the regional total load demand forecast value and the actual value and normalizing it. Furthermore, the action space is defined as the adjustment amount of the charging and discharging power of each energy storage unit. The reward function integrates the economic index (price difference benefit) with the grid stability index (power fluctuation suppression) and the battery aging cost (given by i The charging and discharging power adjustment of energy storage equipment and state of charge (SOC).

[0036] The training process of the deep reinforcement learning model DRL includes state space encoding, action space constraints and training strategy design, such as Figure 2 As shown in the figure, the specific process of DRL training is divided into 5 steps. Environment and agent initialization mainly involves environment configuration (such as defining state space, action space, reward function and termination conditions) and building a neural network; data collection includes experience parameters and interaction processes; network training includes small batch sampling, loss function calculation, and back propagation updates, which can accelerate data collection and cover more state space; strategy evaluation and optimization include exploration strategy and evaluation strategy; convergence judgment and termination include stop conditions and model preservation.

[0037] Specifically, the real-time electricity price signal, the state of charge (SOC) of the energy storage device, the grid frequency regulation demand level, the regional total load demand forecast value, and the load forecast deviation coefficient obtained by calculating the deviation between the regional total load demand forecast value and the actual value and normalizing them are encoded in the state space. Furthermore, the real-time electricity price is normalized to [0,1] (based on the historical highest / lowest price); the SOC state is converted into a 20-dimensional vector (the current SOC value of each energy storage device), and each dimensional vector represents a different attribute of the SOC data.

[0038] The action space constraint is the total power sum( ) The balance formula is: sum(Δ P i ) = P req ±γ%; in, For the i The initial power adjustment of the energy storage equipment, including i =1,...,N, where N is the total number of energy storage devices, P req is the predicted value of regional total load demand, and γ% represents the set deviation. In this embodiment, it is set to 2%.

[0039] The training strategies include distributed parallel training, prioritized experience replay (PER), and training termination condition setting. Distributed parallel training refers to using 32 workers to explore different strategies simultaneously; prioritized experience replay (PER) refers to sampling by temporal difference error (TD) priority; and training termination condition setting refers to reward fluctuations of <1% for 100 consecutive rounds.

[0040] Through the deep reinforcement learning model DRL output i Preliminary power adjustment of energy storage equipment .

[0041] Real-time control is performed on the training process of the deep reinforcement learning model DRL, including operation cycle design, instruction issuance and execution, and exception handling mechanism.

[0042] Specifically, in the long period of minutes, the cascade neural network prediction model is updated and the future load curve with minute granularity is generated; in the short period of seconds, the DRL model generates power allocation instructions according to the latest electricity price and SOC status, and the secondary optimization solver adjusts the initial power adjustment amount of each energy storage device. , satisfying the device constraints, where i=1, ..., N, where N is the total number of energy storage devices. In this embodiment, the long cycle is set to 5 minutes to generate the load curve for the next hour (5-minute granularity); the short cycle is set to 10 seconds.

[0043] The instruction issuance and execution refers to the MQTT protocol (QoS=1, ensuring at least one delivery) used from the cloud to the edge computing node; the communication protocol from the edge computing node to the device uses the CAN bus (transmission cycle 10ms).

[0044] The exception handling mechanism includes SOC balancing trigger conditions and emergency shutdown scenarios.

[0045] Specifically, the SOC balancing trigger condition means that when the maximum SOC difference in the energy storage group is greater than the set SOC difference threshold, the balancing charging mode is started; the maximum SOC difference is the difference between the maximum and minimum SOC values ​​of all energy storage devices in the same energy storage cluster; the balancing charging mode is to set the SOC intermediate value, discharge or charge the energy storage devices above or below the SOC intermediate value respectively, and limit the maximum power of the balancing process to a fixed multiple of the rated capacity of the battery in the energy storage device. In this embodiment, the SOC difference threshold is set to 15%, and the fixed multiple is set to 0.2.

[0046] The emergency shutdown scenario means that when the temperature of the battery in the energy storage device is greater than the set temperature threshold or the voltage change rate in the previous and next two seconds is greater than the change rate threshold, the charging and discharging circuit is cut off, the liquid cooling system is started (flow rate>5L / min), and the cloud is reported to trigger the backup power supply switch. In this embodiment, the set temperature threshold is 65°C and the change rate threshold is set to 10%.

[0047] S4, using the initial power adjustment Construct an objective function that takes into account the battery aging cost, and determine the power adjustment penalty coefficient in the objective function of the optimization model based on the constraints. and the state of charge weight factor The specific steps include: Based on the preliminary power adjustment solution and the Q-learning based quadratic optimization algorithm, an objective function model considering the battery aging cost is established: Aging cost = ; in, is the power adjustment penalty coefficient, For the i The initial power adjustment of the energy storage equipment, including i =1,...,N, where N is the total number of energy storage devices, is the state of charge weight coefficient, SOC i For the iThe state of charge of the energy storage device, SOC ref Indicates the initial capacity or benchmark capacity of an energy storage device under standard test conditions.

[0048] Aging costs need to be kept to a minimum, namely: ; Set up constraints: ; in, SOC i express t +1 moment i The state of charge of the energy storage device, i =1,...,N, where N is the total number of energy storage devices; Indicates i The initial power adjustment of the energy storage equipment; Indicates i The initial power of the energy storage equipment; and Respectively represent i The minimum and maximum values ​​of the power of each energy storage device after adjustment; Represents the total power of N energy storage devices.

[0049] Through the interaction between the Q-learning agent and the environment, the objective function is determined. and The range of parameter values ​​provides direction and balances various optimization requirements. Dynamically adjust weight coefficients and :When the grid frequency regulation demand is high, reduce To allow greater power fluctuations, giving priority to grid stability. When the battery health (SOH) decreases, increase To strengthen SOC balance protection.

[0050] This dynamic adjustment mechanism can flexibly optimize the objective function according to the actual situation, making the power allocation scheme more adaptable. It is the key link to achieve dynamic optimization by further considering the real-time status of the system based on the objective function and constraints.

[0051] S5, power-based penalty coefficient adjustment and the state of charge weight factor The initial value range of and The optimal coefficient solution is obtained, and the final power allocation scheme is determined based on the optimal coefficient solution. The specific steps include: There are multi-objective optimization problems in the process of generating the Pareto optimal solution set, which requires minimizing grid fluctuations, maximizing economic benefits, and minimizing battery aging at the same time, and there are conflicts between these objectives. Therefore, this embodiment proposes a solution step for obtaining the Pareto optimal solution set, including weighted space sampling method, non-dominated sorting (NSGA-II algorithm) and decision maker selection mechanism.

[0052] Specifically, the steps of the weighted space sampling method are: right and Perform grid sampling (e.g. ∈[0.01,0.1], ∈[0.05,0.5]), generate multiple groups of optimization solutions. By sampling the weight coefficients within a certain range, the power allocation schemes under different parameter combinations can be obtained, providing a rich candidate set for subsequent screening of the optimal solution.

[0053] Furthermore, the steps of non-dominated sorting (NSGA-II algorithm) are: The first frontier is determined based on the candidate set, that is, the solution set that is not dominated by other solutions (for example, if the economy and life of solution A are better than solution B, then A dominates B).

[0054] This step selects relatively better solutions from many candidate solutions. These solutions achieve a good balance between multiple objectives and are an important part of the Pareto optimal solution.

[0055] Furthermore, the decision-maker selection mechanism provides two selection methods: economic priority and life priority, and selects the appropriate solution from the Pareto optimal solution set according to actual needs: Economic Priority: Choice Small, Small solution (allowing larger power fluctuations).

[0056] Life span priority: Select big, Large solution (strictly limit SOC and power variation).

[0057] This provides decision makers with flexible options based on different application scenarios and needs, so that the optimization results can better meet the diverse needs in practical applications.

[0058] S6, deploy edge computing nodes to achieve millisecond-level status feedback, monitor the health status of energy storage equipment in real time, and dynamically update the optimization model weight parameters. By real-time monitoring of the health status of the equipment to obtain data, dynamically update the model weight parameters to optimize the charging and discharging control, forming a feedback closed loop. According to the optimal charging and discharging strategy obtained by deep reinforcement learning and cascade neural network decision-making, the battery is controlled in real time through the control system of the energy storage power station to adjust the charging and discharging power and charging and discharging time of each battery. During the control execution process, the operating status and health changes of the battery are monitored in real time, and the actual data is fed back to the equipment health assessment model, multi-objective optimization model and deep reinforcement learning model. According to the feedback information, the model is dynamically adjusted and optimized to achieve real-time closed-loop control and dynamic adjustment of the battery status.

[0059] Specifically, build a device health assessment model: ; in, Indicates i The health scores of energy storage devices are i =1,...,N, where N is the total number of energy storage devices, , 1 indicates the best state; SOH indicates the battery health state, ,in, Indicates the current capacity. Indicates the initial capacity; It represents the temperature influence coefficient, and the calculation formula is: , T is the average battery temperature, T opt For the best working temperature, such as 25℃, k is a constant; Indicates the cumulative number of equivalent cycles; Indicates the battery design life, such as 6000 times; , and They represent the battery health status weight, temperature impact weight and cycle number weight respectively. By setting these weights reasonably, it can ensure that all relevant factors are The contributions of In this embodiment, the value is , , .

[0060] Through the above closed-loop control steps, dynamic charging and discharging control of multi-user energy storage power stations can be achieved, dynamic adjustment and optimization can be performed according to the real-time health status of the battery, health management problems in traditional methods can be solved, and the flexibility and efficiency of charging and discharging control of energy storage power stations can be greatly improved.

[0061] The present invention also proposes a multi-user energy storage power station charging and discharging control system based on electricity price, including a data acquisition and preprocessing module, a load prediction module, a preliminary power adjustment amount calculation module, a power adjustment amount optimization module and a weight parameter update module: The data collection and preprocessing module collects multi-user load data on the user side and performs data preprocessing; The load forecasting module inputs multi-user load data into the cascade neural network forecasting model to generate load demand forecast values ​​for future periods; The preliminary power adjustment calculation module builds a deep reinforcement learning model DRL and outputs the preliminary power adjustment of each energy storage device; The power adjustment optimization module establishes a secondary optimization model that takes into account the battery aging cost and obtains the final power allocation solution that meets the grid demand and equipment constraints; The optimal power allocation scheme design module obtains the optimal coefficient solution based on the preliminary value range and determines the final power allocation scheme; The closed-loop control module calculates the health of the energy storage equipment and performs closed-loop control on the weight parameter update of the power adjustment prediction model and the multi-objective optimization model.

[0062] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

Claims

1. A multi-user energy storage power station charging and discharging control method based on electricity price, characterized in that: include: S1, collects the load data of each energy storage device, including the temperature, voltage and current of the battery, and performs preprocessing; S2, constructing a cascade neural network prediction model, using the preprocessed load data to train the cascade neural network prediction model, and outputting the regional total load demand prediction value; S3, based on the deep reinforcement learning model DRL, builds a power adjustment prediction model. Through the collected real-time electricity price signals, the current state of charge SOC of the energy storage device, the grid frequency regulation demand level, and the regional total load demand forecast value, the training process of the power adjustment prediction model is controlled in real time to obtain the preliminary power adjustment value of each energy storage device. ,in i =1,...,N, where N is the total number of energy storage devices; S4, using the initial power adjustment Construct an objective function that takes into account the battery aging cost, and determine the power adjustment penalty coefficient in the objective function of the optimization model based on the constraints. and the state of charge weight factor The initial value range of ; S5, power-based penalty adjustment and the state of charge weight factor The initial value range of and The optimal coefficient solution is obtained, and the final power allocation plan is determined according to the optimal coefficient solution; the charging and discharging control of the multi-user energy storage power station is carried out based on the power allocation plan.

2. The method for controlling charging and discharging of a multi-user energy storage power station based on electricity price according to claim 1, characterized in that: In S2, the cascade neural network prediction model includes the temporal convolutional network TCN and the graph long short-term memory network Graph-LSTM; Collect user type characteristics and environmental parameters, input historical load data, user type characteristics and environmental parameters into the cascade neural network prediction model, and output the predicted value of regional total load demand.

3. The method for controlling charging and discharging of a multi-user energy storage power station based on electricity price according to claim 1, characterized in that: S3 includes: S301, the state space includes the real-time electricity price signal, the current state of charge SOC of the energy storage device, the grid frequency regulation demand level and the regional total load demand forecast value, and the load forecast deviation coefficient obtained by calculating the deviation between the regional total load demand forecast value and the actual value and normalizing it; the state space is encoded and converted into the corresponding data form; S302, the regional total load demand forecast value with the set deviation is used as the action space constraint, the total power sum( ) The balance formula is: sum(D P i ) = P req ±γ%; in, For the i The initial power adjustment of the energy storage equipment, including i =1,...,N, where N is the total number of energy storage devices, P req is the forecast value of regional total load demand, γ% represents the setting deviation; S303, training strategy design, including distributed parallel training, priority experience replay PER and training termination condition setting.

4. The method for controlling charging and discharging of a multi-user energy storage power station based on electricity price according to claim 3 is characterized in that: In S301, the real-time electricity price signal, the state of charge SOC of the energy storage device, the grid frequency regulation demand level and the regional total load demand forecast value, as well as the load forecast deviation coefficient obtained by calculating the deviation between the regional total load demand forecast value and the actual value, are encoded, and the real-time electricity price is normalized to [0,1]; the current state of charge SOC of each energy storage device is converted into a 20-dimensional vector, each dimensional vector represents a different attribute of the SOC data.

5. The method for controlling charging and discharging of a multi-user energy storage power station based on electricity price according to claim 3 is characterized in that: In S3, the training process of the power adjustment prediction model is controlled in real time, including the operation cycle design: In the long period of minutes, the cascade neural network prediction model is updated and the future load curve with minute granularity is generated; In a short period of seconds, the DRL model generates power allocation instructions based on the latest electricity price and SOC status, and the quadratic optimization solver adjusts the initial power adjustment of each energy storage device. , satisfying the device constraints, where i =1,...,N, where N is the total number of energy storage devices; It also includes instruction issuance and execution and exception handling mechanisms.

6. The method for controlling charging and discharging of a multi-user energy storage power station based on electricity price according to claim 5 is characterized in that: The exception handling mechanism includes SOC balancing trigger conditions and emergency shutdown scenarios; The SOC balancing trigger condition means that when the maximum SOC difference in the energy storage group is greater than the set SOC difference threshold, the balancing charging mode is started; the maximum SOC difference is the difference between the maximum and minimum SOC values ​​of all energy storage devices in the same energy storage cluster; The balanced charging mode is to set the SOC middle value, discharge or charge the energy storage devices above or below the SOC middle value respectively, and limit the maximum power of the balanced process to a fixed multiple of the rated capacity of the battery in the energy storage device; The emergency shutdown scenario means that when the temperature of the battery in the energy storage device is greater than the set temperature threshold or the voltage change rate in the two seconds before and after is greater than the change rate threshold, the charging and discharging circuit is cut off, the liquid cooling system is started, and the cloud is reported to trigger the backup power supply switch.

7. The method for controlling charging and discharging of a multi-user energy storage power station based on electricity price according to claim 1, characterized in that: In S4, the initial power adjustment amount Based on the Q-learning quadratic optimization algorithm, an objective function model considering the battery aging cost is established: Aging cost = ; in, is the power adjustment penalty coefficient, For the i The initial power adjustment of the energy storage equipment, including i =1,...,N, where N is the total number of energy storage devices, is the state of charge weight coefficient, SOC i For the i The state of charge of the energy storage device, SOC ref Indicates the initial capacity or benchmark capacity of the energy storage device under standard test conditions; The aging costs need to be kept to a minimum, namely: 。 8. The method for controlling charging and discharging of a multi-user energy storage power station based on electricity price according to claim 7 is characterized in that: Set up constraints: ； in, SOC i express t +1 moment i The state of charge of the energy storage device, i =1,...,N, where N is the total number of energy storage devices; Indicates i The initial power adjustment of the energy storage equipment; Indicates i The initial power of the energy storage equipment; and Respectively represent i The minimum and maximum values ​​of the power of each energy storage device after adjustment; Represents the total power of N energy storage devices; According to the constraints, the penalty coefficient is adjusted through the interaction between the Q-learning agent and the environment. and the state of charge weight factor The initial value range of .

9. The method for controlling charging and discharging of a multi-user energy storage power station based on electricity price according to claim 1, characterized in that: The multi-objective optimization model considers multi-objective optimization problems, including minimizing grid fluctuations, maximizing economic benefits, and minimizing battery aging. The solution steps of the multi-objective optimization model include weighted space sampling method, non-dominated sorting, and decision maker selection mechanism. The steps of weighted space sampling method are: Penalty factor for power adjustment and the state of charge weight factor Perform grid sampling to obtain power allocation solutions under different parameter combinations and generate candidate sets including multiple groups of optimization solutions; The steps of non-dominated sorting are: Determine a first frontier surface according to the candidate set, that is, a solution set that is not dominated by other solutions; The steps for decision makers to choose a mechanism are: Provides two selection modes: economic priority and life priority, and selects the appropriate solution from the solution set that is not dominated by other solutions according to actual needs.

10. The method for controlling charging and discharging of a multi-user energy storage power station based on electricity price according to claim 1, characterized in that: It also includes S6, calculating the health of the energy storage equipment and performing closed-loop control on the weight parameter update of the power adjustment prediction model and the multi-objective optimization model; the formula of the equipment health evaluation model is as follows: ； in, Indicates i The health scores of energy storage devices are i =1,...,N, where N is the total number of energy storage devices, ; SOH indicates the battery health status, ,in, Indicates the current capacity. Indicates the initial capacity; It represents the temperature influence coefficient, and the calculation formula is: , T is the average battery temperature, T opt For the best working temperature, k is a constant; Indicates the cumulative number of equivalent cycles; Indicates the battery design life; , and They represent the battery health status weight, temperature impact weight and cycle number weight respectively.

11. A multi-user energy storage power station charging and discharging control system based on electricity price, comprising a data acquisition and preprocessing module, a load forecasting module, a preliminary power adjustment amount calculation module, a power adjustment amount optimization module and a weight parameter updating module, characterized in that: The data collection and preprocessing module collects multi-user load data on the user side and performs data preprocessing; The load forecasting module inputs multi-user load data into the cascade neural network forecasting model to generate load demand forecast values ​​for future periods; The preliminary power adjustment calculation module builds a deep reinforcement learning model DRL and outputs the preliminary power adjustment of each energy storage device; The power adjustment optimization module establishes a secondary optimization model that takes into account the battery aging cost and determines the initial value range of the power adjustment penalty coefficient and the state of charge weight coefficient in the objective function of the optimization model; The optimal power allocation scheme design module obtains the optimal coefficient solution based on the preliminary value range and determines the final power allocation scheme; The closed-loop control module calculates the health of the energy storage equipment and performs closed-loop control on the weight parameter update of the power adjustment prediction model and the multi-objective optimization model.

Images